Document cuDNN fallback for deterministic training

[kellyguo11]

Description

Document TORCH_CUDNN_V8_API_DISABLED=1 as an optional setting that may improve run-to-run determinism for convolutional training workloads. The note explains that the setting selects PyTorch's cuDNN v7 API instead of cuDNN v8 execution plans.

No new dependencies are required.

Type of change

• Documentation update

Release backport

• Backport this pull request to the active release branch after it merges into develop

Screenshots

Not applicable.

Checklist

Docker and GPU tests run on demand. Push the commits you want tested, then
comment run-ci on the pull request.

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with uv run isaaclab -f
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works (not applicable to this documentation-only change)
• I have added a changelog fragment under source/<pkg>/changelog.d/ for every touched package (no source package is touched)
• I have added my name to the CONTRIBUTORS.md or my name already exists there

Validation

• uv run --isolated --extra dev --extra ov -- make -C docs current-docs
• uv run isaaclab -f

[isaaclab-review-bot]

Isaac Lab Review Bot

The change adds a narrowly scoped note describing the optional cuDNN v7 fallback for convolutional determinism. The proposed objection that this environment variable became inert in PyTorch 2.3 is not supported by the supplied repository context or evidence, so it does not warrant maintainer action.

• Design and architecture: The note is appropriately placed in the existing PyTorch determinism discussion and leaves the separate rendering and physics determinism responsibilities unchanged.
• API: No Isaac Lab public API, configuration field, CLI argument, dependency, or compatibility surface is changed; the documentation only mentions an external PyTorch environment variable.
• Implementation: The reStructuredText note is correctly formed and clearly presents the setting as optional and workload-specific. No documentation references, includes, or surrounding instructions are disrupted.

No blocking issues. No inline issue met the actionable-evidence threshold; the assessment above records the review feedback.

Automated review; human maintainers own approval decisions.

[greptile-apps]

Greptile Summary

Documents an optional environment setting for convolutional training workloads that may improve run-to-run determinism by selecting PyTorch’s cuDNN v7 API path.

• Adds a reproducibility note for TORCH_CUDNN_V8_API_DISABLED=1.
• Clarifies that the setting should be applied before training starts.

Confidence Score: 5/5

The documentation-only change appears safe to merge.

The new note is narrowly scoped, uses appropriately qualified language, and introduces no accepted correctness, security, or maintainability issue.

Important Files Changed

Filename	Overview
docs/source/features/reproducibility.rst	Adds concise documentation for an optional cuDNN API fallback, with no concrete correctness or documentation issue identified.

_{Reviews (1): Last reviewed commit: "Document cuDNN fallback for deterministi..." | Re-trigger Greptile}

[kellyguo11]

run-ci

[isaaclab-bot]

Backported to release/3.0.0 as 11c0a0d.